We consider a glacial lake as one that formed as a result of modern or ancient glaciation. Contemporary glacial lakes are easily recognized using a combination of glacier inventories and remote sensing images. Ancient glacial lakes can be identified from periglacial geomorphological characteristics, including moraine remnants and U-shaped valleys that are discernible from satellite observations (Post and Mayo, 1971; Westoby et al., 2014; Nie et al., 2018; Martín et al., 2021). A 10 km buffering distance of RGI 6.0 glacier boundaries that has been widely used in previous studies (Zhang et al., 2015; Wang et al., 2020) was created to help map glacial lakes. A few glacial lakes in the study area (a total of 84 lakes for the Sentinel-2 dataset and 55 lakes for the Landsat dataset in 2020) beyond the buffering zone, located near buffering boundaries, were intentionally included due to clear evidence of glaciation (Fig. 3). Landslide-dammed lakes (Chen et al., 2017) in the buffering zone were excluded from our inventories because of their irrelevance to glaciation. All glacial lakes in the study area were mapped according to our definition. We were able to implement this definition by carefully leveraging the spectral properties of glacial lakes and the periglacial geomorphological features that are often evident in remote sensing images (see more in Sect. 4.3 and 4.4).

A human-interactive and semi-automated lake mapping method (Wang et al., 2014; Nie et al., 2017, 2020) was adopted to accurately extract glacial lake extents using Landsat and Sentinel-2 images, based on the Normalized Difference Water Index (NDWI) (Mcfeeters, 1996). The NDWI uses the green and near-infrared bands and is calculated using the following equation: 1NDWI=BandGreen-BandNIRBandGreen+BandNIR, where the green band and near-infrared band were provided by both Landsat and Sentinel multispectral images.

Specifically, the method calculated the NDWI histogram based on the pixels with each user-defined and manually drawn region of interest. The NDWI threshold that separates the lake surface from the land was interactively determined by screening the NDWI histogram against the lake region in the imagery (Wang et al., 2014; Nie et al., 2020). This way, the determined NDWI threshold can be well-tuned to adapt to various spectral conditions of the studied glacial lakes. The raster lake extents segmented by the thresholds were then automatically converted to vector polygons. We first completed the glacial lake inventory in 2020 using this interactive mapping method, and the 2020 inventory was then used as a reference to facilitate the lake mapping for other periods.

The minimum mapping unit (MMU) was set to 5 pixels for both Landsat (0.0045 km2) and Sentinel-2 images (0.0005 km2) in this study. The MMU determines the total number and area of glacial lakes in the dataset and varies in the previous studies, such as 3 pixels (Zhang et al., 2015), 6 pixels (Wang et al., 2020), or 9 pixels (Chen et al., 2021) for a regional scale or 55 pixels (Shugar et al., 2020a) for a global scale. While a smaller threshold leads to a large number of lakes mapped, it also generates larger mapping noises or uncertainties. Considering this signal–noise balance and our focus on identifying prominent glacier–lake dynamics in the study area, we opted to use 5 pixels as the MMU for both Landsat and Sentinel-2 images.

Several procedures were taken to assure the quality assurance and quality control for lake mapping, including (1) visual inspection and modification using the threshold-based mapping method for each lake according to Landsat and Sentinel-2 images and Google Earth at a finer scale overlaying preliminarily lake boundary extraction at the given period; (2) time series check for Landsat-derived glacial lake datasets from 1990 and 2020 and cross-check between Landsat- and Sentinel-2-derived lake datasets in 2020 to reduce errors of omission and commission; (3) topological validation of glacial lake mapping, such as repeated removal and elimination of small sliver polygons; and (4) logical check for lake types between two classification systems of glacial lakes. False lake extents resulting from cloud or snow cover, lake ice, and topographic shadows (Nie et al., 2017, 2020) were modified using the previous semi-automated mapping method based on alternative images acquired in adjacent years. Those procedures were time-consuming but helped to minimize the effect of cloud and snow covers, and lake mapping errors, and to maximize the quality of the produced lake product and the derived glacial lake changes.

Two glacial lake classification systems (GLCSs) have been established based on the relationship of the interaction between glacial lakes and glaciers, as well as lake formation mechanism and dam material properties. In the first GLCS (GLCS1), glacial lakes were classified into four types based on their spatial relationship to upstream glaciers: supraglacial, ice-contact, unconnected-glacier-fed lakes, and non-glacier-fed lakes according to Gardelle et al. (2011) and Carrivick et al. (2013). Alternatively, by combining the formation mechanism of glacial lakes and the properties of natural dam features, glacial lakes were classified into five categories (herein named GLCS2) modified from Yao's classification system (2018): supraglacial, end-moraine-dammed, lateral-moraine-dammed, glacial-erosion lakes, and ice-dammed lakes. Subglacial lakes were excluded due to the mapping challenge from spectral satellite images alone. Characterization and examples of each type are provided in Tables 1 and 2. Individual glacial lakes were categorized into the specific types for each GLCS according to available glacier inventory data, and geomorphological and spectral characteristics were interpreted from Landsat and Sentinel images and Google Earth. The synergy of these two GLCSs is beneficial to predicting glacier–lake evolutions and providing fundamental data for water resource and glacial lake disaster risk assessment.

A total of 18 attribute fields were input into our glacial lake datasets (Table 3). They include lake location (longitude and latitude), lake elevation (centroid elevation), orbital number of the image source, image acquisition date, lake area, lake perimeter, lake types of the two GLCSs, mapping uncertainty, lake water volume and the country, sub-basin, and mountain range associated with the lake. Among the attributes, lake location was calculated based on the centroid of each glacial lake polygon associated with the DEM; N represents northing and E represents easting. The orbital number of the image source was filled with the corresponding satellite image, with the codes expressed as “PxxxRxxx” or “Txxxxx”, where P and R indicate the path and row for Landsat image and T represents the tile of Sentinel-2 image associated with five-digit code of the military grid reference system. SceneID indicated identifying information of the image source for Landsat or Sentinel-2, consisting of the orbital number, sensor ID, and acquisition date (YYYYMMDD) for Landsat image or the orbital number and acquisition date (YYYYMMDD) for Sentinel-2 image. Area and perimeter were automatically calculated based on glacial lake extents. Lake water volume was estimated by an area–volume empirical equation (Cook and Quincey, 2015). Lake types were attributed using the characterization and interpretation marks described in Sect. 4.3. Mapping uncertainty was estimated using our modified equation which will be introduced in Sect. 4.5 and the Appendix tutorial. The located country, sub-basin, and mountain range of each glacial lake were identified by overlapping the geographic boundaries of countries, basins, and mountain ranges.

We mapped 2234 glacial lakes in 2020 across the studied CPEC from Landsat 8 images, with a total area of 86.31±14.98 km2 (Fig. 5a and b). Unconnected-glacier-fed lakes are dominant in the first classification system, followed by non-glacier-fed lakes (Fig. 6), whereas glacial-erosion lakes dominate at both number (1478) and area (57.02 km2) in the second classification system (Fig. 7), followed by end-moraine-dammed lakes and supraglacial lakes. Among the classified lakes, 137 are ice-contact lakes and cover an area of 5.56 km2, implying a higher mean size of ice-contact lakes than supraglacial lakes.

The total number and area of glacial lakes in the study remain relatively stable with a slight increase between 1990 and 2020, and the changes in count and area among various types of glacial lakes vary substantially (Figs. 6 and 7). From 1990 to 2020, the total number of glacial lakes increased by 80 (or 3.70 %), while the area grew by 1.21 km2 (or 1.42 %). In GLCS1, unconnected-glacier-fed lakes have the largest increase in number, followed by ice-contact and non-glacier-fed lakes, whereas supraglacial lakes decreased by 62 in count. Ice-contact lakes expanded by 1.24 km2 (equaling an increase of 26 % in ice-contact lakes), contributing one-third of the total area increase. Supraglacial lakes decreased by 0.85 km2 in area, whereas the areas of unconnected-glacier-fed and non-glacier-fed lakes remained stable as a result of disconnections from glaciers (Fig. 6). In GLCS2, end-moraine-dammed lakes increased by 2.48 km2 and contributed most of the glacial lake area expansion, whereas supraglacial, ice-dammed, and lateral-moraine-dammed lakes decreased slightly in both number and area. Glacial-erosion lakes accounted for the maximum percentage (about 66 % for both count and area) in each period and remained stable (Fig. 7).

Sentinel-derived results show that there are 7560 glacial lakes (103.70±8.45 km2) in 2020 across the entire CPEC under an MMU of 5 pixels (500 m2). Compared with the Landsat-derived product, glacial lakes from Sentinel-2 have similar spatial distribution characteristics (Fig. 5); meanwhile, a larger quantity of glacial lakes, with more accurate boundaries and a greater total lake area, were generated from Sentinel-2 images (Table 4). The smallest size class (0.0005–0.0045 km2) contains the maximum lake number (4969) but the least lake area (7.73±2.62 km2), which is not available in the Landsat-derived lake data due to a coarser spatial resolution. In each size class, the overlap ratios are greater than 85 % in count and area, and there are also a higher number and larger area of glacial lakes from Sentinel images than of Landsat images. Sentinel-2 images (10 m) with a finer spatial resolution produce more glacial lakes than those from Landsat images (30 m). The discrepancy is mainly attributed to the inconsistency of spatial resolutions and image acquisition dates, as discussed in Sect. 6.2.

The uncertainty estimated from our improved equation shows that the relative error of individual glacial lakes decreases when lake size increases or the cell size of remote sensing images reduces (Lyons et al., 2013; Carrivick and Quincey, 2014) (Fig. 8). Total area errors of glacial lakes in the study area are approximately ±14.98 and ±8.45 km2 in 2020 for Landsat and Sentinel-2 datasets, respectively, and the average relative errors are ±17.36 % and ±8.15 %. Generally, small lakes have greater relative errors. For example, the mean relative error is 35.38 % for Landsat-derived glacial lakes between 0.0045 and 0.1 km2 and 10.63 % for glacial lakes greater than 0.1 km2. The mean area error of Sentinel-derived glacial lakes is almost one-third of that extracted from Landsat images for glacial lakes of all or specific size groups. Because the relative error was estimated as a function of satellite image spatial resolution and lake perimeter, the calculated error for a large lake is proportionally smaller than that of a small lake (Salerno et al., 2012), and the error for a Landsat-derived lake is naturally greater than that of a Sentinel-derived lake in the same size group.

Our Landsat- and Sentinel-derived glacial lake dataset match well lake boundaries in Google Earth higher-resolution images (Fig. 9). The mean difference in area is 0.005 km2 between Landsat- and Google Earth-derived lakes and 0.001 km2 between Sentinel- and Google Earth-derived lakes, and major validation samples (84/89) are within the confidence interval of 95 %, indicating high accuracy in lake mapping (Fig. 9c and d). The error of 89 sample lakes is 5.48 % of the total area for Landsat- and Google Earth-derived data and 0.61 % for Sentinel- and Google Earth-derived data. The median (± standard deviation) in a discrepancy of the individual lake area is 7.66±4.96 % for Landsat- and Google Earth-derived data and 4.46±4.62 % for Sentinel- and Google Earth-derived data. Our glacial lake dataset shows satisfactory mapping accuracy, although Sentinel-derived lake data perform more accurately than those from Landsat images. We also validated the sampling of 89 Landsat-derived lakes by the existing Landsat-extracted lake data produced by Wang et al. (2020). A total of 83 lakes are available in Wang's data with a mean difference of 0.005 km2 in the lake area (Fig. A8). This also shows an improvement in our lake product in contrast to the existing dataset.

Glacial lakes from Landsat and Sentinel-2 images have high consistency in number and area with overlap rates from approximately 86 % to 94 % for all lakes greater than 0.0045 km2 (Table 4), indicating a good potential for coordinated utility with Landsat archived observation (Fig. 10). Lake extents extracted from Landsat and Sentinel images match well for various types and sizes (Figs. 10 and 11, Table 4). The best consistency rate reaches 94 % for the glacial lakes between 0.1 and 0.2 km2. The difference in the area of glacial lakes extracted from Landsat and Sentinel-2 images generally lies within the uncertainty ranges.

The spatial resolution of satellite images plays a primary role in the discrepancies in count and area of glacial lakes extracted from Landsat (30 m) and Sentinel-2 (10 m) observations. Due to a finer spatial resolution, Sentinel-2 images can extract more glacial lakes and more accurate extents than those from Landsat images. We set the same 5 pixels as the MMU for both Landsat and Sentinel-2 images, which corresponds to a minimum area of 0.0045 and 0.0005 km2, respectively. The minimum mapping area results in generating nearly 5000 more lakes from Sentinel-2 images than from Landsat images, causing the greatest discrepancy in number, such as Fig. 11. Small lakes such as supraglacial lakes play an important role in analyzing glacier evolution and supraglacial drainage systems (Liu and Mayer, 2015; Miles et al., 2018), implying a potential of our dataset to be applied in studies of glacier–lake evolutions. Meanwhile, Sentinel-2 images can depict boundaries of glacial lakes with lower uncertainty, as some small islands and narrow channels (Fig. 11b and c) were mapped from Sentinel-2 imagery that were unable to be detected in Landsat imagery.

In addition to the difference in image resolution, different acquisition dates between Sentinel-2 and Landsat images can also contribute to the discrepancy between those two glacial lake datasets. The total number of supraglacial lakes and ice-dammed glacial lakes are less than 300, but those lakes are controlled by glacier movement and temperature changes (Liu and Mayer, 2015; Miles et al., 2018), which vary faster with time than relatively stable glacial-erosion and moraine-dammed lakes. Acquiring same-day images from the two sensors was not always possible due to the impacts of cloud contaminations, topographic shadows, snow cover, and revisit periods (Williamson et al., 2018; Paul et al., 2020). Despite our efforts of leveraging all available high-quality images, the overlap of acquisition dates between Landsat and Sentinel-2 images for the same location is relatively low (only 7 scenes of Sentinel-2 images or 112 glacial lakes in 2020) in this study area, and the consequential temporal gaps led to a difference in the number and area of the derived glacial lakes. As exemplified in Fig. 11d, the mapped supraglacial lakes in the same location exhibit a considerable discrepancy, which is likely a joint consequence of both sensor difference and glacial lake evolution.

An increasing number of glacial lake datasets have been released over the past years, and most of them were produced from long-term Landsat archives. Regional glacial lake datasets using Sentinel images are scarce. The lack of Sentinel-derived glacial lake data in the study area makes it impossible to compare. Here we selected four available glacial lake datasets to compare with our Landsat-derived dataset at the same MMU and study area.

We provide the latest glacial lake dataset (in 2020) and the most long-term 30 m Landsat observation (1990 to 2020) for this study, with a range of critical attributes including two types of classification systems. Within the same study area, our 2020 glacial lakes appear to be closest to the 2018 dataset produced by Wang et al. (2020), with the highest overlap of greater than 91 % in count at the MMU of 5400 m2 or 6 pixels (Table 5). Wang's dataset (2020) contains many large landslide-dammed lakes that are excluded in our glacial lake mapping, so their total glacial lake area is greater than ours. The overlapping rates between Wang's glacial lakes (2020) in 1990 and ours are more than 83 % in count. However, their results show a distinct increase of glacial lakes in number and area between 1990 and 2018 (Wang et al., 2020), whereas our data show a more stable change between 1990 and 2020. One possible reason is that manually delineating glacial lakes twice by different operators during Wang's lake mapping (2020) exacerbates the errors of mapping. Another reason is that their data contain landslide-dammed lakes that fluctuate greatly with time and expanded recently. One example is Attabad Lake (located at 36∘18′ 22.33′′ N, 74∘49′ 34.36′′ E).

The second highest overlapping rate is approximately 59 % for 2008 and 58 % for 2017 in count compared with Chen's data (Chen et al., 2021). Similarly, the overlapping rate between Shugar's dataset (2020a) and ours is lower than 43 % in count at the MMU of 50 000 m2. The dataset from Zhang et al. (2015) shows fewer glacial lakes in 1990 and 2000 at the same MMU of 5 pixels. Our product has more lakes than each of the other four products at nine time periods. By inspecting their dataset, we attributed this anomalous discrepancy to a range of glacial lakes that were missing due to a lack of thorough cross-check quality assurance during their lake mapping over a larger study area. And those additional glacial lakes show an improvement of our product in contrast to the previous similar datasets. Our Landsat-derived glacial lake dataset has been visually cross-checked over three time periods after the step of threshold-based semi-automated lake mapping and has also been visually validated by Sentinel-derived glacial lakes. Through this series of quality assurance, we aim at delivering one of the most reliable multi-decadal glacial lake products for this study area.

Other factors, such as image quality and acquisition dates, mapping methods, and a quality assurance workflow, might also lead to discrepancies between the glacial lake datasets. Despite such discrepancies, an increasing number of publicly shared datasets benefit potential users to select the most suitable one for their objectives. Herein, we provide an up-to-date glacial lake dataset derived from both Landsat and Sentinel-2 observations, which further increased the availability of glacial lake datasets for water resource and GLOF risk assessment, predicting glacier–lake evolutions (Carrivick et al., 2020) in the context of climate change.

We would like to acknowledge several limitations of our glacial lake dataset, largely due to the availability of high-quality satellite images in the study area and inadequate field survey data (Wang et al., 2020; Chen et al., 2021). First, it is unlikely for one to collect enough good-quality images within 1 calendar year for the entire study area due to the high possibility of cloud or snow cover. Even though the capacity of repeat observations for Landsat 8 OLI and Sentinel-2 increased (Roy et al., 2014; Williamson et al., 2018; Wulder et al., 2019; Paul et al., 2020), the 2020 glacial lake dataset has to employ images acquired in adjacent years besides 2020. Most images used from Landsat and Sentinel-2 platforms were imaged in autumn, and some images taken between April and July and in November were also employed. Distribution and changes in glacial lakes primarily represent the characteristics between August and October. Glacial lakes evolve with time and space (Nie et al., 2017), and subtle inter- and intra-annual changes (Liu et al., 2020) for each period were ignored. Second, field investigation data are limited due to the low accessibility of the high mountain environment in the study area, which restrained the accuracy in classifying the glacial lake types. Although very high-resolution Google Earth images were utilized to assist in lake-type interpretation, occasional misclassification was unavoidable. We implemented two types of classification systems based on a careful utilization of glacier data, DEM, geomorphological features, and expert knowledge. However, the lack of in situ surveys prohibited a thorough validation of the glacial lake types. Third, the rigorous quality assurance and cross-check after semi-automated lake mapping assure the quality of our lake dataset but are still time- and cost-prohibitive. State-of-the-art mapping methods, such as deep learning (Wu et al., 2020), Google Earth Engine cloud computing (Chen et al., 2021), and synergy of SAR and optical images (Wangchuk and Bolch, 2020; How et al., 2021), could be used in the future to balance product accuracy and time cost.

The glacial lake dataset will be updated using newly collected Landsat and Sentinel images at a 5-year interval or modified according to user feedback. The updated glacial lake dataset will continue to be released freely and publicly on the Mountain Science Data Center sharing platform.